[wieslawsoltes]

Problem

Hover invalidation recompiled every dependency trigger for every changed hover-path node. The work scaled as path depth × dependency count × selector parse cost, so a 4,000-selector workload delayed publication of accepted input.

Change

• compile each hover trigger once while indexing selector dependencies;
• retain its source string for diagnostics and reuse the compiled selector during hover transitions;
• add a focused hover-dependency-cache native runner covering functional/sibling selector behavior and the deep descendant case;
• measure the 4,000-dependency/32-depth transition through the scene carrying the pointer sequence, so the 30 ms gate includes cascade, required layout/tree work, and publication;
• require stable DOM-node count and no more than one layout pass for the transition.

Validation

Focused current head d04ca758702ced1b0e56b8368ff55c9fe0d06da6:

• git diff --check HEAD^ HEAD — pass;
• Release-equivalent -fsyntax-only compile of webscene_v8_runtime.cpp with V8/html5ever/cssparser/Servo/generated-bindings production flags — pass;
• Release-equivalent -fsyntax-only compile of native_v8_runtime_tests.cpp — pass;
• historical uncached control: 72.599 ms;
• original cached candidate: 5.025 ms (14.4× faster).

The current-head runner must pass WEBSCENE_NATIVE_ENGINE_TEST_FILTER=hover-dependency-cache; that result is the authoritative changed-behavior gate. Broad package and product matrices are outside this focused parser/cache change.

Integration

Closes #115.

The equivalent implementation also exists in draft consolidation PR #76. This focused PR targets main directly; #76 remains open and unmerged.